Column-store database architecture utilizing positional delta tree update system and methods

Abstract

A column-store database computer system responsive to database requests for the update and retrieval of data from within a stable data table providing for the storage of database tuples within a column-store organized database structure. A positional delta tree data structure is implemented in the memory space of the computer system and is operatively coupled in an update data transfer path between a database engine interface and the stable data table. The positional delta tree data structure includes a differential data storage layer operative to store differential update data values in positionally defined relative reference to database tuples stored by the stable data table.

Description

[0001]This application claims the benefit of U.S. Provisional Application Nos. 61 / 209,833, 61 / 209,820, and 61 / 209,796, all filed Mar. 11, 2009.LIMITED COPYRIGHT WAIVER[0002]A portion of the disclosure of this patent document contains material to which the claim of copyright protection is made. The copyright owner has no objection to the facsimile reproduction by any parson of the patent document or the patent disclosure as it appears in the US Patent and Trademark Office file or records, but reserves all other rights whatsoever.BACKGROUND OF THE INVENTION[0003]1. Field of the Invention[0004]The present invention is generally related to database storage management systems and, in particular, to a column-store database storage system utilizing a positional delta tree subsystem in high-performance support of database access requests.[0005]2. Description of the Related Art[0006]Column-oriented data base systems, commonly referred to as column-stores have recently regained commercial and...

AI Technical Summary

Benefits of technology

[0017]An advantage of the present invention is that the read performance of a column store database is maintained while enabling fast update operations. Database updates nominally execute against a positional delta tree data structure. Read operations resolve as a merger of data from the underlying column store and the positional delta tree data structure. The positional delta tree structure is specifically designed to minimize the computational and I / O resource requirements for performing on-the-fly data merges.

[0018]Another advantage of the present invention is that the composed structure of the positional delta tree enables the effective pre-computation of read data merge locations, thereby minimizing the I / O operations required to resolve and merge data from the positional delta tree for each database read operation and the corresponding CPU load. Read operation merges using the positional delta tree structure are essentially insensitive to the presence and composition of database table sort keys.

[0019]A further advantage of the present invention is that the performance impact of updates accumulated in a positional delta tree data structure is a substantially linear function of the depth of the positional delta tree data structure. Update operations impose a substantially fixed overhead to pre-compute the position of the update and store the results in the positional delta tree data structure. Read operations incur a scan overhead required to locate any applicable updated data stored in the positional delta tree data structure.

[0020]Still another advantage of the present invention is that the positional delta tree data structure can reduce the requirements for and impacts of database locks on database transactions. Further, the positional delta tree data structure directly supports update isolation in transactions and an efficient conflict check mechanism operative on transaction commit.

[0021]Yet another advantage of the present invention is that the positional delta tree is organized as a functionally layered system, enabling efficient use and management in the implementation of database engine operations, including queries, updates, merges, and checkpoints. A substantial number of updates can be accumulated without significantly affecting query performance. Through systematic management of the positional delta tree data structures, updates captured initially in an in-memory write store positional delta tree layer are subsequently transferred into an in-memory read-store positional delta tree layer, and then to the underlying column store data tables as stored in a persistent read-store. The transfers preferably occur in response to merge and checkpoint events, and complete with little significant impact on ongoing query performance. The layered, consistently managed positional delta tree data structures thus makes the database system generally insensitive to whether updates are performed as individual updates or batched as bulk updates.

[0022]Still a further advantage of the present invention is that the positional delta tree data structure is specifically designed to be memory space efficient. While portions may be copied to a persistent data store, preferably the entire structure resides in and operates from within a combination of the CPU cache memory and main memory store of a database computer system. By maintaining the positional delta tree data structure in memory, on-the-fly merger of update and read data is achieved without requiring additional persistent data store I / O operations. By allowing specific portions of the positional delta tree data structure to be loaded and maintained in CPU cache memory, the number of main memory store accesses is reduced, resulting in a significant reduction of overall processing overhead and improved performance particularly in completing on-the-fly merger computations.

Problems solved by technology

The cumulative disk utilization of update operations also creates the potential for blocking reads, thereby directly impacting the primary operation of the column-store.

The complexity and overhead of read operations are, however, substantially increased by the necessity to perform on-the-fly merger of tuples as read separately from the read and write-stores.

While the LSMT represents the current conventionally preferred write-store structure for column-stores, the read performance penalty associated with use of the LSMT remains a substantial impediment to the practical adoption and use of column-stores.

This required expansive read-scan and related data merge directly imposes substantial time and resource expensive disk I / O to retrieve sort key attribute values as well as significant CPU overhead due to the complexity of arbitrary data type, multicolumn merge operations, resulting in degradation of all column-store related operations.

Images